11CS 594 Spring 2003Lecture 1:Overview of High-Performance ComputingJack Dongarra Computer Science DepartmentUniversity of Tennessee 2Most Important Slideu Netlib - software repositoryØGo to http://www.netlib.org/u Register for the na-digestØGo to http://www.netlib.org/na-net/ØRegister to receive the na-digest» http://www.netlib.org/na-net/join_mail_forw.html3Computational Scienceu HPC offers a new way to do science:Ø Experiment - Theory - Computationu Computation used to approximate physical systems - Advantages include:ØPlaying with simulation parameters to study emergent trendsØPossible replay of a particular simulation eventØStudy systems where no exact theories exist 4Why Turn to Simulation?u When the problem is too . . .Ø ComplexØ Large / smallØ ExpensiveØ Dangerousu to do any other way.Taurus_to_Taurus_60per_30deg.mpeg5Pretty Pictures6Automotive Industryu Huge users of HPC technology;Ø Ford is 25th largest user of HPC in the worldu Main uses of simulation:Ø Aerodynamics (similar to aerospace)Ø Crash simulationØ Metal sheet formationØ Noise/vibration optimizationØ Traffic simulationu Main benefits:Ø Reduced time to market of new carsØ Increased qualityØ Reduced need to build prototypesØ more efficient & integrated manufacturing processes27Why Turn to Simulation?u Climate / Weather Modelingu Data intensive problems (data-mining, oil reservoir simulation)u Problems with large length and time scales (cosmology)8Units of High Performance Computing1 Mflop/s 1 Megaflop/s 106 Flop/sec1 Gflop/s 1 Gigaflop/s 109 Flop/sec1 Tflop/s 1 Teraflop/s 1012 Flop/sec1 Pflop/s 1 Petaflop/s 1015 Flop/sec1 MB 1 Megabyte 106 Bytes1 GB 1 Gigabyte 109 Bytes1 TB 1 Terabyte 1012 Bytes1 PB 1 Petabyte 1015 Bytes9High-Performance Computing Todayu In the past decade, the world has experienced one of the most exciting periods in computer development.u Microprocessors have become smaller, denser, and more powerful.u The result is that microprocessor-based supercomputing is rapidly becoming the technology of preference in attacking some of the most important problems of science and engineering.10Technology Trends: Microprocessor Capacity2X transistors/Chip Every 1.5 yearsCalled “Moore’s Law”Moore’s LawMicroprocessors have become smaller, denser, and more powerful.Not just processors, bandwidth, storage, etcGordon Moore (co-founder of Intel) predicted in 1965 that the transistor density of semiconductor chips would double roughly every 18 months. 11Internet –4thRevolution in Telecommunicationsu Telephone, Radio, Televisionu Growth in Internet outstrips the othersu Exponential growth since 1985u Traffic doubles every 100 days12The Web Phenomenonu 90 – 93 Web inventedu U of Illinois Mosaic released March 94, ~ 0.1% trafficu September 93 ~ 1% traffic w/200 sitesu June 94 ~ 10% of traffic w/2,000 sitesu Today 60% of traffic w/2,000,000 sitesu Every organization, company, school313Peer to Peer Computingu Peer-to-peer is a style of networking in which a group of computers communicate directly with each other.u Wireless communicationu Home computer in the utility room, next to the water heater and furnace. u Web tablets u Imbedded computers in things all tied together.Ø Books, furniture, milk cartons, etcu Smart AppliancesØ Refrigerator, scale, etc14Internet On Everything15SETI@home: Global Distributed Computingu Running on 500,000 PCs, ~1000 CPU Years per DayØ 485,821 CPU Years so faru Sophisticated Data & Signal Processing Analysisu Distributes Datasets from Arecibo Radio Telescope16SETI@homeu Use thousands of Internet-connected PCs to help in the search for extraterrestrial intelligence.u When their computer is idle or being wasted this software will download a 300 kilobyte chunk of data for analysis. Performs about 3 Tflops for each client in 15 hours.u The results of this analysis are sent back to the SETI team, combined with thousands of other participants.u Largest distributed computation project in existenceØ Averaging 40 Tflop/su Today a number of companies trying this for profit.17Grid Computing -from ET toAnthrax18u Google query attributesØ 150M queries/day (2000/second)Ø 3B documents in the indexu Data centersØ 15,000 Linux systems in 6 data centers» 15 TFlop/s and 1000 TB total capability» 40-80 1U/2U servers/cabinet » 100 MB Ethernet switches/cabinet with gigabit Ethernet uplinkØ growth from 4,000 systems (June 2000)» 18M queries thenu Performance and operationØ simple reissue of failed commands to new serversØ no performance debugging » problems are not reproducibleSource: Monika Henzinger, Google419Next Generation Webu To treat CPU cycles and software like commodities.u Enable the coordinated use of geographically distributed resources – in the absence of central control and existing trust relationships. u Computing power is produced much like utilities such as power and water are produced for consumers.u Users will have access to “power” on demand u This is one of our efforts at UT.Performance vs. TimeMips25 mhz0.11.010100Performance (VAX 780s)1980 19851990MV10K68K7805 MhzRISC 60% / yruVAX 6K(CMOS)8600TTLECL 15%/yrCMOS CISC 38%/yro ||MIPS (8 Mhz)o9000Mips(65 Mhz)uVAXCMOSWill RISC continue on a 60%, (x4 / 3 years)? Moore's speed law?4K21Trends in Computer PerformanceoASCI Red22Other Examples: Sony PlayStation2u Emotion Engine: 6.2 Gflop/s, 75 million polygons per second (Microprocessor Report, 13:5)Ø Superscalar MIPS core + vector coprocessor + graphics/DRAMØ Claim: “Toy Story” realism brought to gamesØ About $25023Sony PlayStation2 Export Limits?24Where Has This Performance Improvement Come From?u Technology?u Organization?u Instruction Set Architecture?u Software?u Some combination of all of the above?51st PrinciplesuWhat happens when the feature size shrinks by a factor of x ?uClock rate goes up by xØactually less than x, because of power consumptionuTransistors per unit area goes up by x2uDie size also tends to increaseØtypically another factor of ~xuRaw computing power of the chip goes up by ~ x4 !Øof which x3 is devoted either to parallelism or locality26How fast can a serial computer be?u Consider the 1 Tflop sequential machineØ data must travel some distance, r, to get from memory to CPUØ to get 1 data element per cycle, this means 1012 times per second at the speed of light, c = 3x108m/sØ so r < c/1012= .3 mmu Now put 1 TB of storage in a .3 mm2area Ø each word occupies about 3 Angstroms2, the size of
View Full Document
Unlocking...